GLM 5.2 and the coming AI margin collapse (part 1)

Everyone remembers the DeepSeek panic, when markets decided cheap training meant the AI capex boom was over. That was the wrong lesson. Training is a one-time cost you amortize; inference is where the real money is made, and where the real margins hide. Anthropic and OpenAI charging $25 per million tokens looks a lot less like cost recovery and more like a business built on selling very profitable API calls once the training bill is paid off.

GLM5.2 from Z.ai is the model that finally makes that math uncomfortable for the frontier labs. After weeks of daily use, the author found it genuinely hard to distinguish from Opus in real coding sessions. It thinks a lot, which slows it down for interactive work and isn't ideal, and it has no vision support and weak web search, both real gaps against Claude and GPT. But for background agentic tasks like reviewing pull requests, none of that matters much.

What should worry Anthropic and OpenAI more is how frictionless switching has become. Both Z.ai and Fireworks expose OpenAI- and Anthropic-compatible endpoints, so pointing Claude Code or Codex at GLM5.2 is a one-line config change, not a migration project. There's no vendor lock-in here comparable to swapping out Salesforce. Enterprise data privacy concerns around Z.ai's own China-linked hosting are real, but plenty of third-party providers offer proper contracts, and self-hosting is an option too.

Then there's price. GLM5.2 runs around $4.40 per million tokens, less than a fifth of Opus's retail rate and about 15% of GPT5.5's. It burns more tokens per task since it thinks more, so it's not a perfectly clean comparison, but even accounting for that it's likely over 50% cheaper for most real workflows at comparable quality. And costs are still falling: Wafer's benchmarking found GLM5.2 running 2.75x cheaper per token on AMD hardware versus Nvidia Blackwell, with more serving-stack optimization clearly still to come.

None of this kills the frontier labs outright, but it puts real pressure on the fat margins that made their business model work so cleanly. When switching costs approach zero and quality gaps approach zero, pricing power evaporates fast.